1 // SPDX-License-Identifier: MIT
2 /*
3 * Copyright © 2020 Intel Corporation
4 */
5
6 #include <linux/log2.h>
7
8 #include "gem/i915_gem_internal.h"
9 #include "gem/i915_gem_lmem.h"
10
11 #include "gen8_ppgtt.h"
12 #include "i915_scatterlist.h"
13 #include "i915_trace.h"
14 #include "i915_pvinfo.h"
15 #include "i915_vgpu.h"
16 #include "intel_gt.h"
17 #include "intel_gtt.h"
18
gen8_pde_encode(const dma_addr_t addr,const enum i915_cache_level level)19 static u64 gen8_pde_encode(const dma_addr_t addr,
20 const enum i915_cache_level level)
21 {
22 u64 pde = addr | GEN8_PAGE_PRESENT | GEN8_PAGE_RW;
23
24 if (level != I915_CACHE_NONE)
25 pde |= PPAT_CACHED_PDE;
26 else
27 pde |= PPAT_UNCACHED;
28
29 return pde;
30 }
31
gen8_pte_encode(dma_addr_t addr,unsigned int pat_index,u32 flags)32 static u64 gen8_pte_encode(dma_addr_t addr,
33 unsigned int pat_index,
34 u32 flags)
35 {
36 gen8_pte_t pte = addr | GEN8_PAGE_PRESENT | GEN8_PAGE_RW;
37
38 if (unlikely(flags & PTE_READ_ONLY))
39 pte &= ~GEN8_PAGE_RW;
40
41 /*
42 * For pre-gen12 platforms pat_index is the same as enum
43 * i915_cache_level, so the switch-case here is still valid.
44 * See translation table defined by LEGACY_CACHELEVEL.
45 */
46 switch (pat_index) {
47 case I915_CACHE_NONE:
48 pte |= PPAT_UNCACHED;
49 break;
50 case I915_CACHE_WT:
51 pte |= PPAT_DISPLAY_ELLC;
52 break;
53 default:
54 pte |= PPAT_CACHED;
55 break;
56 }
57
58 return pte;
59 }
60
gen12_pte_encode(dma_addr_t addr,unsigned int pat_index,u32 flags)61 static u64 gen12_pte_encode(dma_addr_t addr,
62 unsigned int pat_index,
63 u32 flags)
64 {
65 gen8_pte_t pte = addr | GEN8_PAGE_PRESENT | GEN8_PAGE_RW;
66
67 if (unlikely(flags & PTE_READ_ONLY))
68 pte &= ~GEN8_PAGE_RW;
69
70 if (flags & PTE_LM)
71 pte |= GEN12_PPGTT_PTE_LM;
72
73 if (pat_index & BIT(0))
74 pte |= GEN12_PPGTT_PTE_PAT0;
75
76 if (pat_index & BIT(1))
77 pte |= GEN12_PPGTT_PTE_PAT1;
78
79 if (pat_index & BIT(2))
80 pte |= GEN12_PPGTT_PTE_PAT2;
81
82 if (pat_index & BIT(3))
83 pte |= MTL_PPGTT_PTE_PAT3;
84
85 return pte;
86 }
87
gen8_ppgtt_notify_vgt(struct i915_ppgtt * ppgtt,bool create)88 static void gen8_ppgtt_notify_vgt(struct i915_ppgtt *ppgtt, bool create)
89 {
90 struct drm_i915_private *i915 = ppgtt->vm.i915;
91 struct intel_uncore *uncore = ppgtt->vm.gt->uncore;
92 enum vgt_g2v_type msg;
93 int i;
94
95 if (create)
96 atomic_inc(px_used(ppgtt->pd)); /* never remove */
97 else
98 atomic_dec(px_used(ppgtt->pd));
99
100 mutex_lock(&i915->vgpu.lock);
101
102 if (i915_vm_is_4lvl(&ppgtt->vm)) {
103 const u64 daddr = px_dma(ppgtt->pd);
104
105 intel_uncore_write(uncore,
106 vgtif_reg(pdp[0].lo), lower_32_bits(daddr));
107 intel_uncore_write(uncore,
108 vgtif_reg(pdp[0].hi), upper_32_bits(daddr));
109
110 msg = create ?
111 VGT_G2V_PPGTT_L4_PAGE_TABLE_CREATE :
112 VGT_G2V_PPGTT_L4_PAGE_TABLE_DESTROY;
113 } else {
114 for (i = 0; i < GEN8_3LVL_PDPES; i++) {
115 const u64 daddr = i915_page_dir_dma_addr(ppgtt, i);
116
117 intel_uncore_write(uncore,
118 vgtif_reg(pdp[i].lo),
119 lower_32_bits(daddr));
120 intel_uncore_write(uncore,
121 vgtif_reg(pdp[i].hi),
122 upper_32_bits(daddr));
123 }
124
125 msg = create ?
126 VGT_G2V_PPGTT_L3_PAGE_TABLE_CREATE :
127 VGT_G2V_PPGTT_L3_PAGE_TABLE_DESTROY;
128 }
129
130 /* g2v_notify atomically (via hv trap) consumes the message packet. */
131 intel_uncore_write(uncore, vgtif_reg(g2v_notify), msg);
132
133 mutex_unlock(&i915->vgpu.lock);
134 }
135
136 /* Index shifts into the pagetable are offset by GEN8_PTE_SHIFT [12] */
137 #define GEN8_PAGE_SIZE (SZ_4K) /* page and page-directory sizes are the same */
138 #define GEN8_PTE_SHIFT (ilog2(GEN8_PAGE_SIZE))
139 #define GEN8_PDES (GEN8_PAGE_SIZE / sizeof(u64))
140 #define gen8_pd_shift(lvl) ((lvl) * ilog2(GEN8_PDES))
141 #define gen8_pd_index(i, lvl) i915_pde_index((i), gen8_pd_shift(lvl))
142 #define __gen8_pte_shift(lvl) (GEN8_PTE_SHIFT + gen8_pd_shift(lvl))
143 #define __gen8_pte_index(a, lvl) i915_pde_index((a), __gen8_pte_shift(lvl))
144
145 #define as_pd(x) container_of((x), typeof(struct i915_page_directory), pt)
146
147 static unsigned int
gen8_pd_range(u64 start,u64 end,int lvl,unsigned int * idx)148 gen8_pd_range(u64 start, u64 end, int lvl, unsigned int *idx)
149 {
150 const int shift = gen8_pd_shift(lvl);
151 const u64 mask = ~0ull << gen8_pd_shift(lvl + 1);
152
153 GEM_BUG_ON(start >= end);
154 end += ~mask >> gen8_pd_shift(1);
155
156 *idx = i915_pde_index(start, shift);
157 if ((start ^ end) & mask)
158 return GEN8_PDES - *idx;
159 else
160 return i915_pde_index(end, shift) - *idx;
161 }
162
gen8_pd_contains(u64 start,u64 end,int lvl)163 static bool gen8_pd_contains(u64 start, u64 end, int lvl)
164 {
165 const u64 mask = ~0ull << gen8_pd_shift(lvl + 1);
166
167 GEM_BUG_ON(start >= end);
168 return (start ^ end) & mask && (start & ~mask) == 0;
169 }
170
gen8_pt_count(u64 start,u64 end)171 static unsigned int gen8_pt_count(u64 start, u64 end)
172 {
173 GEM_BUG_ON(start >= end);
174 if ((start ^ end) >> gen8_pd_shift(1))
175 return GEN8_PDES - (start & (GEN8_PDES - 1));
176 else
177 return end - start;
178 }
179
gen8_pd_top_count(const struct i915_address_space * vm)180 static unsigned int gen8_pd_top_count(const struct i915_address_space *vm)
181 {
182 unsigned int shift = __gen8_pte_shift(vm->top);
183
184 return (vm->total + (1ull << shift) - 1) >> shift;
185 }
186
187 static struct i915_page_directory *
gen8_pdp_for_page_index(struct i915_address_space * const vm,const u64 idx)188 gen8_pdp_for_page_index(struct i915_address_space * const vm, const u64 idx)
189 {
190 struct i915_ppgtt * const ppgtt = i915_vm_to_ppgtt(vm);
191
192 if (vm->top == 2)
193 return ppgtt->pd;
194 else
195 return i915_pd_entry(ppgtt->pd, gen8_pd_index(idx, vm->top));
196 }
197
198 static struct i915_page_directory *
gen8_pdp_for_page_address(struct i915_address_space * const vm,const u64 addr)199 gen8_pdp_for_page_address(struct i915_address_space * const vm, const u64 addr)
200 {
201 return gen8_pdp_for_page_index(vm, addr >> GEN8_PTE_SHIFT);
202 }
203
__gen8_ppgtt_cleanup(struct i915_address_space * vm,struct i915_page_directory * pd,int count,int lvl)204 static void __gen8_ppgtt_cleanup(struct i915_address_space *vm,
205 struct i915_page_directory *pd,
206 int count, int lvl)
207 {
208 if (lvl) {
209 void **pde = pd->entry;
210
211 do {
212 if (!*pde)
213 continue;
214
215 __gen8_ppgtt_cleanup(vm, *pde, GEN8_PDES, lvl - 1);
216 } while (pde++, --count);
217 }
218
219 free_px(vm, &pd->pt, lvl);
220 }
221
gen8_ppgtt_cleanup(struct i915_address_space * vm)222 static void gen8_ppgtt_cleanup(struct i915_address_space *vm)
223 {
224 struct i915_ppgtt *ppgtt = i915_vm_to_ppgtt(vm);
225
226 if (vm->rsvd.obj)
227 i915_gem_object_put(vm->rsvd.obj);
228
229 if (intel_vgpu_active(vm->i915))
230 gen8_ppgtt_notify_vgt(ppgtt, false);
231
232 if (ppgtt->pd)
233 __gen8_ppgtt_cleanup(vm, ppgtt->pd,
234 gen8_pd_top_count(vm), vm->top);
235
236 free_scratch(vm);
237 }
238
__gen8_ppgtt_clear(struct i915_address_space * const vm,struct i915_page_directory * const pd,u64 start,const u64 end,int lvl)239 static u64 __gen8_ppgtt_clear(struct i915_address_space * const vm,
240 struct i915_page_directory * const pd,
241 u64 start, const u64 end, int lvl)
242 {
243 const struct drm_i915_gem_object * const scratch = vm->scratch[lvl];
244 unsigned int idx, len;
245
246 GEM_BUG_ON(end > vm->total >> GEN8_PTE_SHIFT);
247
248 len = gen8_pd_range(start, end, lvl--, &idx);
249 GTT_TRACE("%s(%p):{ lvl:%d, start:%llx, end:%llx, idx:%d, len:%d, used:%d }\n",
250 __func__, vm, lvl + 1, start, end,
251 idx, len, atomic_read(px_used(pd)));
252 GEM_BUG_ON(!len || len >= atomic_read(px_used(pd)));
253
254 do {
255 struct i915_page_table *pt = pd->entry[idx];
256
257 if (atomic_fetch_inc(&pt->used) >> gen8_pd_shift(1) &&
258 gen8_pd_contains(start, end, lvl)) {
259 GTT_TRACE("%s(%p):{ lvl:%d, idx:%d, start:%llx, end:%llx } removing pd\n",
260 __func__, vm, lvl + 1, idx, start, end);
261 clear_pd_entry(pd, idx, scratch);
262 __gen8_ppgtt_cleanup(vm, as_pd(pt), I915_PDES, lvl);
263 start += (u64)I915_PDES << gen8_pd_shift(lvl);
264 continue;
265 }
266
267 if (lvl) {
268 start = __gen8_ppgtt_clear(vm, as_pd(pt),
269 start, end, lvl);
270 } else {
271 unsigned int count;
272 unsigned int pte = gen8_pd_index(start, 0);
273 unsigned int num_ptes;
274 u64 *vaddr;
275
276 count = gen8_pt_count(start, end);
277 GTT_TRACE("%s(%p):{ lvl:%d, start:%llx, end:%llx, idx:%d, len:%d, used:%d } removing pte\n",
278 __func__, vm, lvl, start, end,
279 gen8_pd_index(start, 0), count,
280 atomic_read(&pt->used));
281 GEM_BUG_ON(!count || count >= atomic_read(&pt->used));
282
283 num_ptes = count;
284 if (pt->is_compact) {
285 GEM_BUG_ON(num_ptes % 16);
286 GEM_BUG_ON(pte % 16);
287 num_ptes /= 16;
288 pte /= 16;
289 }
290
291 vaddr = px_vaddr(pt);
292 memset64(vaddr + pte,
293 vm->scratch[0]->encode,
294 num_ptes);
295
296 atomic_sub(count, &pt->used);
297 start += count;
298 }
299
300 if (release_pd_entry(pd, idx, pt, scratch))
301 free_px(vm, pt, lvl);
302 } while (idx++, --len);
303
304 return start;
305 }
306
gen8_ppgtt_clear(struct i915_address_space * vm,u64 start,u64 length)307 static void gen8_ppgtt_clear(struct i915_address_space *vm,
308 u64 start, u64 length)
309 {
310 GEM_BUG_ON(!IS_ALIGNED(start, BIT_ULL(GEN8_PTE_SHIFT)));
311 GEM_BUG_ON(!IS_ALIGNED(length, BIT_ULL(GEN8_PTE_SHIFT)));
312 GEM_BUG_ON(range_overflows(start, length, vm->total));
313
314 start >>= GEN8_PTE_SHIFT;
315 length >>= GEN8_PTE_SHIFT;
316 GEM_BUG_ON(length == 0);
317
318 __gen8_ppgtt_clear(vm, i915_vm_to_ppgtt(vm)->pd,
319 start, start + length, vm->top);
320 }
321
__gen8_ppgtt_alloc(struct i915_address_space * const vm,struct i915_vm_pt_stash * stash,struct i915_page_directory * const pd,u64 * const start,const u64 end,int lvl)322 static void __gen8_ppgtt_alloc(struct i915_address_space * const vm,
323 struct i915_vm_pt_stash *stash,
324 struct i915_page_directory * const pd,
325 u64 * const start, const u64 end, int lvl)
326 {
327 unsigned int idx, len;
328
329 GEM_BUG_ON(end > vm->total >> GEN8_PTE_SHIFT);
330
331 len = gen8_pd_range(*start, end, lvl--, &idx);
332 GTT_TRACE("%s(%p):{ lvl:%d, start:%llx, end:%llx, idx:%d, len:%d, used:%d }\n",
333 __func__, vm, lvl + 1, *start, end,
334 idx, len, atomic_read(px_used(pd)));
335 GEM_BUG_ON(!len || (idx + len - 1) >> gen8_pd_shift(1));
336
337 spin_lock(&pd->lock);
338 GEM_BUG_ON(!atomic_read(px_used(pd))); /* Must be pinned! */
339 do {
340 struct i915_page_table *pt = pd->entry[idx];
341
342 if (!pt) {
343 spin_unlock(&pd->lock);
344
345 GTT_TRACE("%s(%p):{ lvl:%d, idx:%d } allocating new tree\n",
346 __func__, vm, lvl + 1, idx);
347
348 pt = stash->pt[!!lvl];
349 __i915_gem_object_pin_pages(pt->base);
350
351 fill_px(pt, vm->scratch[lvl]->encode);
352
353 spin_lock(&pd->lock);
354 if (likely(!pd->entry[idx])) {
355 stash->pt[!!lvl] = pt->stash;
356 atomic_set(&pt->used, 0);
357 set_pd_entry(pd, idx, pt);
358 } else {
359 pt = pd->entry[idx];
360 }
361 }
362
363 if (lvl) {
364 atomic_inc(&pt->used);
365 spin_unlock(&pd->lock);
366
367 __gen8_ppgtt_alloc(vm, stash,
368 as_pd(pt), start, end, lvl);
369
370 spin_lock(&pd->lock);
371 atomic_dec(&pt->used);
372 GEM_BUG_ON(!atomic_read(&pt->used));
373 } else {
374 unsigned int count = gen8_pt_count(*start, end);
375
376 GTT_TRACE("%s(%p):{ lvl:%d, start:%llx, end:%llx, idx:%d, len:%d, used:%d } inserting pte\n",
377 __func__, vm, lvl, *start, end,
378 gen8_pd_index(*start, 0), count,
379 atomic_read(&pt->used));
380
381 atomic_add(count, &pt->used);
382 /* All other pdes may be simultaneously removed */
383 GEM_BUG_ON(atomic_read(&pt->used) > NALLOC * I915_PDES);
384 *start += count;
385 }
386 } while (idx++, --len);
387 spin_unlock(&pd->lock);
388 }
389
gen8_ppgtt_alloc(struct i915_address_space * vm,struct i915_vm_pt_stash * stash,u64 start,u64 length)390 static void gen8_ppgtt_alloc(struct i915_address_space *vm,
391 struct i915_vm_pt_stash *stash,
392 u64 start, u64 length)
393 {
394 GEM_BUG_ON(!IS_ALIGNED(start, BIT_ULL(GEN8_PTE_SHIFT)));
395 GEM_BUG_ON(!IS_ALIGNED(length, BIT_ULL(GEN8_PTE_SHIFT)));
396 GEM_BUG_ON(range_overflows(start, length, vm->total));
397
398 start >>= GEN8_PTE_SHIFT;
399 length >>= GEN8_PTE_SHIFT;
400 GEM_BUG_ON(length == 0);
401
402 __gen8_ppgtt_alloc(vm, stash, i915_vm_to_ppgtt(vm)->pd,
403 &start, start + length, vm->top);
404 }
405
__gen8_ppgtt_foreach(struct i915_address_space * vm,struct i915_page_directory * pd,u64 * start,u64 end,int lvl,void (* fn)(struct i915_address_space * vm,struct i915_page_table * pt,void * data),void * data)406 static void __gen8_ppgtt_foreach(struct i915_address_space *vm,
407 struct i915_page_directory *pd,
408 u64 *start, u64 end, int lvl,
409 void (*fn)(struct i915_address_space *vm,
410 struct i915_page_table *pt,
411 void *data),
412 void *data)
413 {
414 unsigned int idx, len;
415
416 len = gen8_pd_range(*start, end, lvl--, &idx);
417
418 spin_lock(&pd->lock);
419 do {
420 struct i915_page_table *pt = pd->entry[idx];
421
422 atomic_inc(&pt->used);
423 spin_unlock(&pd->lock);
424
425 if (lvl) {
426 __gen8_ppgtt_foreach(vm, as_pd(pt), start, end, lvl,
427 fn, data);
428 } else {
429 fn(vm, pt, data);
430 *start += gen8_pt_count(*start, end);
431 }
432
433 spin_lock(&pd->lock);
434 atomic_dec(&pt->used);
435 } while (idx++, --len);
436 spin_unlock(&pd->lock);
437 }
438
gen8_ppgtt_foreach(struct i915_address_space * vm,u64 start,u64 length,void (* fn)(struct i915_address_space * vm,struct i915_page_table * pt,void * data),void * data)439 static void gen8_ppgtt_foreach(struct i915_address_space *vm,
440 u64 start, u64 length,
441 void (*fn)(struct i915_address_space *vm,
442 struct i915_page_table *pt,
443 void *data),
444 void *data)
445 {
446 start >>= GEN8_PTE_SHIFT;
447 length >>= GEN8_PTE_SHIFT;
448
449 __gen8_ppgtt_foreach(vm, i915_vm_to_ppgtt(vm)->pd,
450 &start, start + length, vm->top,
451 fn, data);
452 }
453
454 static __always_inline u64
gen8_ppgtt_insert_pte(struct i915_ppgtt * ppgtt,struct i915_page_directory * pdp,struct sgt_dma * iter,u64 idx,unsigned int pat_index,u32 flags)455 gen8_ppgtt_insert_pte(struct i915_ppgtt *ppgtt,
456 struct i915_page_directory *pdp,
457 struct sgt_dma *iter,
458 u64 idx,
459 unsigned int pat_index,
460 u32 flags)
461 {
462 struct i915_page_directory *pd;
463 const gen8_pte_t pte_encode = ppgtt->vm.pte_encode(0, pat_index, flags);
464 gen8_pte_t *vaddr;
465
466 pd = i915_pd_entry(pdp, gen8_pd_index(idx, 2));
467 vaddr = px_vaddr(i915_pt_entry(pd, gen8_pd_index(idx, 1)));
468 do {
469 GEM_BUG_ON(sg_dma_len(iter->sg) < I915_GTT_PAGE_SIZE);
470 vaddr[gen8_pd_index(idx, 0)] = pte_encode | iter->dma;
471
472 iter->dma += I915_GTT_PAGE_SIZE;
473 if (iter->dma >= iter->max) {
474 iter->sg = __sg_next(iter->sg);
475 if (!iter->sg || sg_dma_len(iter->sg) == 0) {
476 idx = 0;
477 break;
478 }
479
480 iter->dma = sg_dma_address(iter->sg);
481 iter->max = iter->dma + sg_dma_len(iter->sg);
482 }
483
484 if (gen8_pd_index(++idx, 0) == 0) {
485 if (gen8_pd_index(idx, 1) == 0) {
486 /* Limited by sg length for 3lvl */
487 if (gen8_pd_index(idx, 2) == 0)
488 break;
489
490 pd = pdp->entry[gen8_pd_index(idx, 2)];
491 }
492
493 drm_clflush_virt_range(vaddr, PAGE_SIZE);
494 vaddr = px_vaddr(i915_pt_entry(pd, gen8_pd_index(idx, 1)));
495 }
496 } while (1);
497 drm_clflush_virt_range(vaddr, PAGE_SIZE);
498
499 return idx;
500 }
501
502 static void
xehp_ppgtt_insert_huge(struct i915_address_space * vm,struct i915_vma_resource * vma_res,struct sgt_dma * iter,unsigned int pat_index,u32 flags)503 xehp_ppgtt_insert_huge(struct i915_address_space *vm,
504 struct i915_vma_resource *vma_res,
505 struct sgt_dma *iter,
506 unsigned int pat_index,
507 u32 flags)
508 {
509 const gen8_pte_t pte_encode = vm->pte_encode(0, pat_index, flags);
510 unsigned int rem = sg_dma_len(iter->sg);
511 u64 start = vma_res->start;
512 u64 end = start + vma_res->vma_size;
513
514 GEM_BUG_ON(!i915_vm_is_4lvl(vm));
515
516 do {
517 struct i915_page_directory * const pdp =
518 gen8_pdp_for_page_address(vm, start);
519 struct i915_page_directory * const pd =
520 i915_pd_entry(pdp, __gen8_pte_index(start, 2));
521 struct i915_page_table *pt =
522 i915_pt_entry(pd, __gen8_pte_index(start, 1));
523 gen8_pte_t encode = pte_encode;
524 unsigned int page_size;
525 gen8_pte_t *vaddr;
526 u16 index, max, nent, i;
527
528 max = I915_PDES;
529 nent = 1;
530
531 if (vma_res->bi.page_sizes.sg & I915_GTT_PAGE_SIZE_2M &&
532 IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_2M) &&
533 rem >= I915_GTT_PAGE_SIZE_2M &&
534 !__gen8_pte_index(start, 0)) {
535 index = __gen8_pte_index(start, 1);
536 encode |= GEN8_PDE_PS_2M;
537 page_size = I915_GTT_PAGE_SIZE_2M;
538
539 vaddr = px_vaddr(pd);
540 } else {
541 index = __gen8_pte_index(start, 0);
542 page_size = I915_GTT_PAGE_SIZE;
543
544 if (vma_res->bi.page_sizes.sg & I915_GTT_PAGE_SIZE_64K) {
545 /*
546 * Device local-memory on these platforms should
547 * always use 64K pages or larger (including GTT
548 * alignment), therefore if we know the whole
549 * page-table needs to be filled we can always
550 * safely use the compact-layout. Otherwise fall
551 * back to the TLB hint with PS64. If this is
552 * system memory we only bother with PS64.
553 */
554 if ((encode & GEN12_PPGTT_PTE_LM) &&
555 end - start >= SZ_2M && !index) {
556 index = __gen8_pte_index(start, 0) / 16;
557 page_size = I915_GTT_PAGE_SIZE_64K;
558
559 max /= 16;
560
561 vaddr = px_vaddr(pd);
562 vaddr[__gen8_pte_index(start, 1)] |= GEN12_PDE_64K;
563
564 pt->is_compact = true;
565 } else if (IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_64K) &&
566 rem >= I915_GTT_PAGE_SIZE_64K &&
567 !(index % 16)) {
568 encode |= GEN12_PTE_PS64;
569 page_size = I915_GTT_PAGE_SIZE_64K;
570 nent = 16;
571 }
572 }
573
574 vaddr = px_vaddr(pt);
575 }
576
577 do {
578 GEM_BUG_ON(rem < page_size);
579
580 for (i = 0; i < nent; i++) {
581 vaddr[index++] =
582 encode | (iter->dma + i *
583 I915_GTT_PAGE_SIZE);
584 }
585
586 start += page_size;
587 iter->dma += page_size;
588 rem -= page_size;
589 if (iter->dma >= iter->max) {
590 iter->sg = __sg_next(iter->sg);
591 if (!iter->sg)
592 break;
593
594 rem = sg_dma_len(iter->sg);
595 if (!rem)
596 break;
597
598 iter->dma = sg_dma_address(iter->sg);
599 iter->max = iter->dma + rem;
600
601 if (unlikely(!IS_ALIGNED(iter->dma, page_size)))
602 break;
603 }
604 } while (rem >= page_size && index < max);
605
606 drm_clflush_virt_range(vaddr, PAGE_SIZE);
607 vma_res->page_sizes_gtt |= page_size;
608 } while (iter->sg && sg_dma_len(iter->sg));
609 }
610
gen8_ppgtt_insert_huge(struct i915_address_space * vm,struct i915_vma_resource * vma_res,struct sgt_dma * iter,unsigned int pat_index,u32 flags)611 static void gen8_ppgtt_insert_huge(struct i915_address_space *vm,
612 struct i915_vma_resource *vma_res,
613 struct sgt_dma *iter,
614 unsigned int pat_index,
615 u32 flags)
616 {
617 const gen8_pte_t pte_encode = vm->pte_encode(0, pat_index, flags);
618 unsigned int rem = sg_dma_len(iter->sg);
619 u64 start = vma_res->start;
620
621 GEM_BUG_ON(!i915_vm_is_4lvl(vm));
622
623 do {
624 struct i915_page_directory * const pdp =
625 gen8_pdp_for_page_address(vm, start);
626 struct i915_page_directory * const pd =
627 i915_pd_entry(pdp, __gen8_pte_index(start, 2));
628 gen8_pte_t encode = pte_encode;
629 unsigned int maybe_64K = -1;
630 unsigned int page_size;
631 gen8_pte_t *vaddr;
632 u16 index;
633
634 if (vma_res->bi.page_sizes.sg & I915_GTT_PAGE_SIZE_2M &&
635 IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_2M) &&
636 rem >= I915_GTT_PAGE_SIZE_2M &&
637 !__gen8_pte_index(start, 0)) {
638 index = __gen8_pte_index(start, 1);
639 encode |= GEN8_PDE_PS_2M;
640 page_size = I915_GTT_PAGE_SIZE_2M;
641
642 vaddr = px_vaddr(pd);
643 } else {
644 struct i915_page_table *pt =
645 i915_pt_entry(pd, __gen8_pte_index(start, 1));
646
647 index = __gen8_pte_index(start, 0);
648 page_size = I915_GTT_PAGE_SIZE;
649
650 if (!index &&
651 vma_res->bi.page_sizes.sg & I915_GTT_PAGE_SIZE_64K &&
652 IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_64K) &&
653 (IS_ALIGNED(rem, I915_GTT_PAGE_SIZE_64K) ||
654 rem >= (I915_PDES - index) * I915_GTT_PAGE_SIZE))
655 maybe_64K = __gen8_pte_index(start, 1);
656
657 vaddr = px_vaddr(pt);
658 }
659
660 do {
661 GEM_BUG_ON(sg_dma_len(iter->sg) < page_size);
662 vaddr[index++] = encode | iter->dma;
663
664 start += page_size;
665 iter->dma += page_size;
666 rem -= page_size;
667 if (iter->dma >= iter->max) {
668 iter->sg = __sg_next(iter->sg);
669 if (!iter->sg)
670 break;
671
672 rem = sg_dma_len(iter->sg);
673 if (!rem)
674 break;
675
676 iter->dma = sg_dma_address(iter->sg);
677 iter->max = iter->dma + rem;
678
679 if (maybe_64K != -1 && index < I915_PDES &&
680 !(IS_ALIGNED(iter->dma, I915_GTT_PAGE_SIZE_64K) &&
681 (IS_ALIGNED(rem, I915_GTT_PAGE_SIZE_64K) ||
682 rem >= (I915_PDES - index) * I915_GTT_PAGE_SIZE)))
683 maybe_64K = -1;
684
685 if (unlikely(!IS_ALIGNED(iter->dma, page_size)))
686 break;
687 }
688 } while (rem >= page_size && index < I915_PDES);
689
690 drm_clflush_virt_range(vaddr, PAGE_SIZE);
691
692 /*
693 * Is it safe to mark the 2M block as 64K? -- Either we have
694 * filled whole page-table with 64K entries, or filled part of
695 * it and have reached the end of the sg table and we have
696 * enough padding.
697 */
698 if (maybe_64K != -1 &&
699 (index == I915_PDES ||
700 (i915_vm_has_scratch_64K(vm) &&
701 !iter->sg && IS_ALIGNED(vma_res->start +
702 vma_res->node_size,
703 I915_GTT_PAGE_SIZE_2M)))) {
704 vaddr = px_vaddr(pd);
705 vaddr[maybe_64K] |= GEN8_PDE_IPS_64K;
706 drm_clflush_virt_range(vaddr, PAGE_SIZE);
707 page_size = I915_GTT_PAGE_SIZE_64K;
708
709 /*
710 * We write all 4K page entries, even when using 64K
711 * pages. In order to verify that the HW isn't cheating
712 * by using the 4K PTE instead of the 64K PTE, we want
713 * to remove all the surplus entries. If the HW skipped
714 * the 64K PTE, it will read/write into the scratch page
715 * instead - which we detect as missing results during
716 * selftests.
717 */
718 if (I915_SELFTEST_ONLY(vm->scrub_64K)) {
719 u16 i;
720
721 encode = vm->scratch[0]->encode;
722 vaddr = px_vaddr(i915_pt_entry(pd, maybe_64K));
723
724 for (i = 1; i < index; i += 16)
725 memset64(vaddr + i, encode, 15);
726
727 drm_clflush_virt_range(vaddr, PAGE_SIZE);
728 }
729 }
730
731 vma_res->page_sizes_gtt |= page_size;
732 } while (iter->sg && sg_dma_len(iter->sg));
733 }
734
gen8_ppgtt_insert(struct i915_address_space * vm,struct i915_vma_resource * vma_res,unsigned int pat_index,u32 flags)735 static void gen8_ppgtt_insert(struct i915_address_space *vm,
736 struct i915_vma_resource *vma_res,
737 unsigned int pat_index,
738 u32 flags)
739 {
740 struct i915_ppgtt * const ppgtt = i915_vm_to_ppgtt(vm);
741 struct sgt_dma iter = sgt_dma(vma_res);
742
743 if (vma_res->bi.page_sizes.sg > I915_GTT_PAGE_SIZE) {
744 if (GRAPHICS_VER_FULL(vm->i915) >= IP_VER(12, 55))
745 xehp_ppgtt_insert_huge(vm, vma_res, &iter, pat_index, flags);
746 else
747 gen8_ppgtt_insert_huge(vm, vma_res, &iter, pat_index, flags);
748 } else {
749 u64 idx = vma_res->start >> GEN8_PTE_SHIFT;
750
751 do {
752 struct i915_page_directory * const pdp =
753 gen8_pdp_for_page_index(vm, idx);
754
755 idx = gen8_ppgtt_insert_pte(ppgtt, pdp, &iter, idx,
756 pat_index, flags);
757 } while (idx);
758
759 vma_res->page_sizes_gtt = I915_GTT_PAGE_SIZE;
760 }
761 }
762
gen8_ppgtt_insert_entry(struct i915_address_space * vm,dma_addr_t addr,u64 offset,unsigned int pat_index,u32 flags)763 static void gen8_ppgtt_insert_entry(struct i915_address_space *vm,
764 dma_addr_t addr,
765 u64 offset,
766 unsigned int pat_index,
767 u32 flags)
768 {
769 u64 idx = offset >> GEN8_PTE_SHIFT;
770 struct i915_page_directory * const pdp =
771 gen8_pdp_for_page_index(vm, idx);
772 struct i915_page_directory *pd =
773 i915_pd_entry(pdp, gen8_pd_index(idx, 2));
774 struct i915_page_table *pt = i915_pt_entry(pd, gen8_pd_index(idx, 1));
775 gen8_pte_t *vaddr;
776
777 GEM_BUG_ON(pt->is_compact);
778
779 vaddr = px_vaddr(pt);
780 vaddr[gen8_pd_index(idx, 0)] = vm->pte_encode(addr, pat_index, flags);
781 drm_clflush_virt_range(&vaddr[gen8_pd_index(idx, 0)], sizeof(*vaddr));
782 }
783
xehp_ppgtt_insert_entry_lm(struct i915_address_space * vm,dma_addr_t addr,u64 offset,unsigned int pat_index,u32 flags)784 static void xehp_ppgtt_insert_entry_lm(struct i915_address_space *vm,
785 dma_addr_t addr,
786 u64 offset,
787 unsigned int pat_index,
788 u32 flags)
789 {
790 u64 idx = offset >> GEN8_PTE_SHIFT;
791 struct i915_page_directory * const pdp =
792 gen8_pdp_for_page_index(vm, idx);
793 struct i915_page_directory *pd =
794 i915_pd_entry(pdp, gen8_pd_index(idx, 2));
795 struct i915_page_table *pt = i915_pt_entry(pd, gen8_pd_index(idx, 1));
796 gen8_pte_t *vaddr;
797
798 GEM_BUG_ON(!IS_ALIGNED(addr, SZ_64K));
799 GEM_BUG_ON(!IS_ALIGNED(offset, SZ_64K));
800
801 /* XXX: we don't strictly need to use this layout */
802
803 if (!pt->is_compact) {
804 vaddr = px_vaddr(pd);
805 vaddr[gen8_pd_index(idx, 1)] |= GEN12_PDE_64K;
806 pt->is_compact = true;
807 }
808
809 vaddr = px_vaddr(pt);
810 vaddr[gen8_pd_index(idx, 0) / 16] = vm->pte_encode(addr, pat_index, flags);
811 }
812
xehp_ppgtt_insert_entry(struct i915_address_space * vm,dma_addr_t addr,u64 offset,unsigned int pat_index,u32 flags)813 static void xehp_ppgtt_insert_entry(struct i915_address_space *vm,
814 dma_addr_t addr,
815 u64 offset,
816 unsigned int pat_index,
817 u32 flags)
818 {
819 if (flags & PTE_LM)
820 return xehp_ppgtt_insert_entry_lm(vm, addr, offset,
821 pat_index, flags);
822
823 return gen8_ppgtt_insert_entry(vm, addr, offset, pat_index, flags);
824 }
825
gen8_init_scratch(struct i915_address_space * vm)826 static int gen8_init_scratch(struct i915_address_space *vm)
827 {
828 u32 pte_flags;
829 int ret;
830 int i;
831
832 /*
833 * If everybody agrees to not to write into the scratch page,
834 * we can reuse it for all vm, keeping contexts and processes separate.
835 */
836 if (vm->has_read_only && vm->gt->vm && !i915_is_ggtt(vm->gt->vm)) {
837 struct i915_address_space *clone = vm->gt->vm;
838
839 GEM_BUG_ON(!clone->has_read_only);
840
841 vm->scratch_order = clone->scratch_order;
842 for (i = 0; i <= vm->top; i++)
843 vm->scratch[i] = i915_gem_object_get(clone->scratch[i]);
844
845 return 0;
846 }
847
848 ret = setup_scratch_page(vm);
849 if (ret)
850 return ret;
851
852 pte_flags = vm->has_read_only;
853 if (i915_gem_object_is_lmem(vm->scratch[0]))
854 pte_flags |= PTE_LM;
855
856 vm->scratch[0]->encode =
857 vm->pte_encode(px_dma(vm->scratch[0]),
858 i915_gem_get_pat_index(vm->i915,
859 I915_CACHE_NONE),
860 pte_flags);
861
862 for (i = 1; i <= vm->top; i++) {
863 struct drm_i915_gem_object *obj;
864
865 obj = vm->alloc_pt_dma(vm, I915_GTT_PAGE_SIZE_4K);
866 if (IS_ERR(obj)) {
867 ret = PTR_ERR(obj);
868 goto free_scratch;
869 }
870
871 ret = map_pt_dma(vm, obj);
872 if (ret) {
873 i915_gem_object_put(obj);
874 goto free_scratch;
875 }
876
877 fill_px(obj, vm->scratch[i - 1]->encode);
878 obj->encode = gen8_pde_encode(px_dma(obj), I915_CACHE_NONE);
879
880 vm->scratch[i] = obj;
881 }
882
883 return 0;
884
885 free_scratch:
886 while (i--)
887 i915_gem_object_put(vm->scratch[i]);
888 vm->scratch[0] = NULL;
889 return ret;
890 }
891
gen8_preallocate_top_level_pdp(struct i915_ppgtt * ppgtt)892 static int gen8_preallocate_top_level_pdp(struct i915_ppgtt *ppgtt)
893 {
894 struct i915_address_space *vm = &ppgtt->vm;
895 struct i915_page_directory *pd = ppgtt->pd;
896 unsigned int idx;
897
898 GEM_BUG_ON(vm->top != 2);
899 GEM_BUG_ON(gen8_pd_top_count(vm) != GEN8_3LVL_PDPES);
900
901 for (idx = 0; idx < GEN8_3LVL_PDPES; idx++) {
902 struct i915_page_directory *pde;
903 int err;
904
905 pde = alloc_pd(vm);
906 if (IS_ERR(pde))
907 return PTR_ERR(pde);
908
909 err = map_pt_dma(vm, pde->pt.base);
910 if (err) {
911 free_pd(vm, pde);
912 return err;
913 }
914
915 fill_px(pde, vm->scratch[1]->encode);
916 set_pd_entry(pd, idx, pde);
917 atomic_inc(px_used(pde)); /* keep pinned */
918 }
919 wmb();
920
921 return 0;
922 }
923
924 static struct i915_page_directory *
gen8_alloc_top_pd(struct i915_address_space * vm)925 gen8_alloc_top_pd(struct i915_address_space *vm)
926 {
927 const unsigned int count = gen8_pd_top_count(vm);
928 struct i915_page_directory *pd;
929 int err;
930
931 GEM_BUG_ON(count > I915_PDES);
932
933 pd = __alloc_pd(count);
934 if (unlikely(!pd))
935 return ERR_PTR(-ENOMEM);
936
937 pd->pt.base = vm->alloc_pt_dma(vm, I915_GTT_PAGE_SIZE_4K);
938 if (IS_ERR(pd->pt.base)) {
939 err = PTR_ERR(pd->pt.base);
940 pd->pt.base = NULL;
941 goto err_pd;
942 }
943
944 err = map_pt_dma(vm, pd->pt.base);
945 if (err)
946 goto err_pd;
947
948 fill_page_dma(px_base(pd), vm->scratch[vm->top]->encode, count);
949 atomic_inc(px_used(pd)); /* mark as pinned */
950 return pd;
951
952 err_pd:
953 free_pd(vm, pd);
954 return ERR_PTR(err);
955 }
956
gen8_init_rsvd(struct i915_address_space * vm)957 static int gen8_init_rsvd(struct i915_address_space *vm)
958 {
959 struct drm_i915_private *i915 = vm->i915;
960 struct drm_i915_gem_object *obj;
961 struct i915_vma *vma;
962 int ret;
963
964 if (!intel_gt_needs_wa_16018031267(vm->gt))
965 return 0;
966
967 /* The memory will be used only by GPU. */
968 obj = i915_gem_object_create_lmem(i915, PAGE_SIZE,
969 I915_BO_ALLOC_VOLATILE |
970 I915_BO_ALLOC_GPU_ONLY);
971 if (IS_ERR(obj))
972 obj = i915_gem_object_create_internal(i915, PAGE_SIZE);
973 if (IS_ERR(obj))
974 return PTR_ERR(obj);
975
976 vma = i915_vma_instance(obj, vm, NULL);
977 if (IS_ERR(vma)) {
978 ret = PTR_ERR(vma);
979 goto unref;
980 }
981
982 ret = i915_vma_pin(vma, 0, 0, PIN_USER | PIN_HIGH);
983 if (ret)
984 goto unref;
985
986 vm->rsvd.vma = i915_vma_make_unshrinkable(vma);
987 vm->rsvd.obj = obj;
988 vm->total -= vma->node.size;
989 return 0;
990 unref:
991 i915_gem_object_put(obj);
992 return ret;
993 }
994
995 /*
996 * GEN8 legacy ppgtt programming is accomplished through a max 4 PDP registers
997 * with a net effect resembling a 2-level page table in normal x86 terms. Each
998 * PDP represents 1GB of memory 4 * 512 * 512 * 4096 = 4GB legacy 32b address
999 * space.
1000 *
1001 */
gen8_ppgtt_create(struct intel_gt * gt,unsigned long lmem_pt_obj_flags)1002 struct i915_ppgtt *gen8_ppgtt_create(struct intel_gt *gt,
1003 unsigned long lmem_pt_obj_flags)
1004 {
1005 struct i915_page_directory *pd;
1006 struct i915_ppgtt *ppgtt;
1007 int err;
1008
1009 ppgtt = kzalloc(sizeof(*ppgtt), GFP_KERNEL);
1010 if (!ppgtt)
1011 return ERR_PTR(-ENOMEM);
1012
1013 ppgtt_init(ppgtt, gt, lmem_pt_obj_flags);
1014 ppgtt->vm.top = i915_vm_is_4lvl(&ppgtt->vm) ? 3 : 2;
1015 ppgtt->vm.pd_shift = ilog2(SZ_4K * SZ_4K / sizeof(gen8_pte_t));
1016
1017 /*
1018 * From bdw, there is hw support for read-only pages in the PPGTT.
1019 *
1020 * Gen11 has HSDES#:1807136187 unresolved. Disable ro support
1021 * for now.
1022 *
1023 * Gen12 has inherited the same read-only fault issue from gen11.
1024 */
1025 ppgtt->vm.has_read_only = !IS_GRAPHICS_VER(gt->i915, 11, 12);
1026
1027 if (HAS_LMEM(gt->i915))
1028 ppgtt->vm.alloc_pt_dma = alloc_pt_lmem;
1029 else
1030 ppgtt->vm.alloc_pt_dma = alloc_pt_dma;
1031
1032 /*
1033 * Using SMEM here instead of LMEM has the advantage of not reserving
1034 * high performance memory for a "never" used filler page. It also
1035 * removes the device access that would be required to initialise the
1036 * scratch page, reducing pressure on an even scarcer resource.
1037 */
1038 ppgtt->vm.alloc_scratch_dma = alloc_pt_dma;
1039
1040 if (GRAPHICS_VER(gt->i915) >= 12)
1041 ppgtt->vm.pte_encode = gen12_pte_encode;
1042 else
1043 ppgtt->vm.pte_encode = gen8_pte_encode;
1044
1045 ppgtt->vm.bind_async_flags = I915_VMA_LOCAL_BIND;
1046 ppgtt->vm.insert_entries = gen8_ppgtt_insert;
1047 if (HAS_64K_PAGES(gt->i915))
1048 ppgtt->vm.insert_page = xehp_ppgtt_insert_entry;
1049 else
1050 ppgtt->vm.insert_page = gen8_ppgtt_insert_entry;
1051 ppgtt->vm.allocate_va_range = gen8_ppgtt_alloc;
1052 ppgtt->vm.clear_range = gen8_ppgtt_clear;
1053 ppgtt->vm.foreach = gen8_ppgtt_foreach;
1054 ppgtt->vm.cleanup = gen8_ppgtt_cleanup;
1055
1056 err = gen8_init_scratch(&ppgtt->vm);
1057 if (err)
1058 goto err_put;
1059
1060 pd = gen8_alloc_top_pd(&ppgtt->vm);
1061 if (IS_ERR(pd)) {
1062 err = PTR_ERR(pd);
1063 goto err_put;
1064 }
1065 ppgtt->pd = pd;
1066
1067 if (!i915_vm_is_4lvl(&ppgtt->vm)) {
1068 err = gen8_preallocate_top_level_pdp(ppgtt);
1069 if (err)
1070 goto err_put;
1071 }
1072
1073 if (intel_vgpu_active(gt->i915))
1074 gen8_ppgtt_notify_vgt(ppgtt, true);
1075
1076 err = gen8_init_rsvd(&ppgtt->vm);
1077 if (err)
1078 goto err_put;
1079
1080 return ppgtt;
1081
1082 err_put:
1083 i915_vm_put(&ppgtt->vm);
1084 return ERR_PTR(err);
1085 }
1086